Skip to main content

Gradium

Since v1.112.0

Gradium is a voice AI provider. LiteLLM calls Gradium's text-to-speech API through the OpenAI-compatible /audio/speech endpoint, so existing OpenAI clients work without code changes.

PropertyDetails
DescriptionConvert text to natural-sounding speech using Gradium's voice AI models.
Provider Route on LiteLLMgradium/
Provider DocGradium TTS ↗
Supported Endpoints/audio/speech

Models​

Use gradium/default for Gradium's default model, or gradium/gradium-tts-beta for the beta model. Any other name after gradium/ is forwarded to Gradium as its model_name. The cost map has prices for gradium/default and gradium/gradium-tts-beta.

Credentials and regions​

LiteLLM reads no Gradium environment variable and has no default region. Every deployment or call must set both api_key and api_base. The key is sent upstream as the x-api-key header.

API baseRegion
https://api.gradium.aiClosest region
https://eu.api.gradium.aiEU
https://us.api.gradium.aiUS

The trailing /api that Gradium's own docs show on these hosts is optional here, so https://eu.api.gradium.ai and https://eu.api.gradium.ai/api reach the same endpoint. When you add a Gradium model with the Add Model form in the LiteLLM UI, the API Base field is a dropdown with these three bases.

Quick Start​

LiteLLM Python SDK​

Generate speech with the SDK
import os
from pathlib import Path

import litellm

response = litellm.speech(
model="gradium/default",
input="the quick brown fox jumped over the lazy dogs",
voice="alloy",
response_format="wav",
api_key=os.environ["GRADIUM_API_KEY"],
api_base="https://eu.api.gradium.ai",
)
response.write_to_file(Path("speech.wav"))

litellm.aspeech takes the same arguments for async code.

LiteLLM Proxy​

1. Configure your proxy​

config.yaml
model_list:
- model_name: gradium-tts
litellm_params:
model: gradium/default
api_key: os.environ/GRADIUM_API_KEY
api_base: https://eu.api.gradium.ai

general_settings:
master_key: os.environ/LITELLM_MASTER_KEY

os.environ/GRADIUM_API_KEY tells the proxy to read the key from that variable when it loads the config. You can pick any variable name.

2. Start the proxy​

Start LiteLLM proxy server
export GRADIUM_API_KEY="your-gradium-api-key"
export LITELLM_MASTER_KEY="sk-your-master-key"
litellm --config config.yaml

3. Generate speech​

curl
curl http://localhost:4000/v1/audio/speech \
-H "Authorization: Bearer $LITELLM_MASTER_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gradium-tts",
"input": "the quick brown fox jumped over the lazy dogs",
"voice": "alloy",
"response_format": "wav"
}' \
--output speech.wav

Streaming​

Set stream: true in the request body, or stream=True in the SDK, to forward audio to the client as Gradium sends it. Without it, LiteLLM buffers the whole clip and returns it in one response.

Voices​

The OpenAI voice names alloy, ash, ballad, coral, echo, fable, nova, onyx, sage, shimmer, verse, marin and cedar map to English voices in Gradium's catalog, picked to match the OpenAI voice's gender. Any other value is sent to Gradium as a raw voice ID, so you can pass the ID of any Gradium voice, including a voice in another language or a custom or cloned one, as voice. For French text, pick a French voice from Gradium's voice library and pass its ID. A request without a voice answers 400.

Response formats​

response_formatGradium output_format
wav (default)wav
pcmpcm_24000 (24 kHz)
opusopus

Gradium's own format names are also accepted and sent as is. These are ulaw_8000, mulaw_8000, alaw_8000, pcm_8000, pcm_16000, pcm_22050, pcm_24000, pcm_44100 and pcm_48000.

Unsupported parameters​

Gradium has no mp3, aac or flac encoder, so those formats answer 400 instead of returning audio with a misleading format. instructions, and any speed other than OpenAI's default of 1.0, have no equivalent on Gradium's endpoint, so they answer 400 as well. To change the speaking pace, use Gradium's own padding_bonus setting, described below.

Set drop_params to drop these instead. A dropped format falls back to wav. You can set it on the call (drop_params=True in the SDK, or "drop_params": true in the request body), on the deployment (drop_params: true under litellm_params), or for every model (litellm.drop_params = True, or drop_params: true under litellm_settings in the proxy config).

Gradium settings​

Send Gradium's json_config in the request body and LiteLLM forwards it to Gradium unchanged. For example, padding_bonus sets the pace. A negative value speaks faster and a positive value speaks slower, as described in Gradium's voice settings.

Speak faster
curl http://localhost:4000/v1/audio/speech \
-H "Authorization: Bearer $LITELLM_MASTER_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gradium-tts",
"input": "the quick brown fox jumped over the lazy dogs",
"voice": "alloy",
"response_format": "wav",
"json_config": {"padding_bonus": -2.0}
}' \
--output speech.wav

In the SDK, pass json_config={"padding_bonus": -2.0} to litellm.speech. With the OpenAI client, pass it in extra_body.

Cost tracking​

Gradium bills 1 credit per input character, spaces included, and LiteLLM bills the same character count. The cost map prices gradium/default and gradium/gradium-tts-beta at the add-on credit rate of the XS plan, $7 per 100k credits, which is $0.00007 per character.

Larger plans are cheaper per credit, and Pro Voice Clone voices bill 1.2 credits per character. LiteLLM bills both at the XS rate for now, because a deployment's input_cost_per_character does not yet change the cost of speech calls. Gradium's response also does not say which kind of voice it used, so LiteLLM cannot tell a clone voice from a catalog voice.

LiteLLM Enterprise
SSO/SAML, audit logs, spend tracking, multi-team management, and guardrails, built for production.
Learn more →